Does convolution in ArrayFire use FFT?

Question

I am looking forward to speed up convolution with derivative of gaussian kernels (upto order 2/3) on large medical images (512 x 512 x 1000 double) in one of our open-source toolkits. We currently do this convolution via FFT.

After being suggested by a friend about ArrayFire and after reading this post, I am trying to see if I could adopt this toolkit. Seems like a great effort and enables us to handle multiple backends though I am currently interested in CUDA alone as that's what I have in hand.

I read this post on the forum that says that convolution in ArrayFire switches to frequency domain after a particular kernel size. I looked at the cuda file convolve.cu but I didn't find any calls to fft within ArrayFire or any of cuFFT stuff. Am i missing something?

Going forward, I would like to construct the derivative of Gaussian kernel directly in frequency domain, multiply with image FFT and bring it back. But I would like to compare the speed ups between creating the convolution kernel in space vs freq domain. Also, ArrayFire doesn't seem to have a Gaussian kernel in 3D.

Answer 1

Your observation is correct, we haven't implemented the fft fallback for larger kernel sizes yet. However, we are trying out best to add this functionality in the following couple of weeks. What kernel sizes are you using in your application ?

We currently do not have 3D Gaussian kernel function implemented. But i believe it can be easily implemented using the JIT enabled element wise operations in ArrayFire. For an example on how to use element wise operations, please look here.